-
Contributions to the continuity problem for
Lyapunov exponents
Adriana Sánchez
August, 2018
-
Contributions to the continuity problem for
Lyapunov exponents
Adriana Sánchez
1. Reviewer Carlos Bocker NetoDepartamento de Matemática
Universidade Federal da Paraíba
2. Reviewer Karina MarínDepartamento de Matemática
Universidade Federal de Minas Gerais
3. Reviewer Lorenzo J. DiazDepartamento de Matemática
Pontificia Universidade Católica do Rio de Janeiro
4. Reviewer Enrique PujalsInstituto de Matemática Pura e Aplicada
5. Reviewer Jacob PalisInstituto de Matemática Pura e Aplicada
Advisor Marcelo Viana
August, 2018
-
Adriana Sánchez
Contributions to the continuity problem for Lyapunov exponents
August, 2018
Reviewers: Carlos Bocker Neto , Karina Marín , Lorenzo J. Diaz , Enrique Pujals and Jacob
Palis
Advisor: Marcelo Viana
Instituto de Matemática Pura e Aplicada
Estrada Dona Castorina 110, Jardim Botânico.
Rio de Janeiro
22460-320
-
Abstract
The aim of this work is to study the continuity and semi-continuity of the Lyapunov
exponents in two different contexts. The first one concerns linear cocycles on par-
tially hyperbolic dynamics. It is known that the Lyapunov exponents can be very
sensitive as functions of the cocycle. Example of this is the result of Bochi-Mañé
which shows that every SL(2,R)-cocycle that is not uniformly hyperbolic can be ap-
proximated by another with zero exponents. We prove that the set of fiber-bunched
SL(2,R)-valued Hölder cocycles with nonvanishing Lyapunov exponents over a vol-
ume preserving, accessible and center-bunched partially hyperbolic diffeomorphism
is open. Moreover, we present an example showing that this is no longer true if we
do not assume accessibility in the base dynamics. This is a joint work with Lucas
Backes and Mauricio Poletti.
In the second part of this work we will restrict our attention to the study of llocally
constant cocycles associated with probability distributions with non-compact sup-
port in SL(2,R). Bocker-Viana proved that for distributions with compact support,
the exponents vary continuously. We analyze the behavior of the Lyapunov expo-
nents when the measures are not compact, showing that in this case, the Lyapunov
exponents, considered as functions of the measure, are semi-continuous with re-
spect to the Wasserstein topology but not the weak* topology. Moreover, we prove
that they are not continuous relative to the Wasserstein topology.
v
-
Resumo
O objetivo deste trabalho é estudar a continuidade e a semi-continuidade dos ex-
poentes de Lyapunov em dois contextos diferentes. O primeiro diz respeito a cocic-
los lineares sobre dinâmicas parcialmente hiperbólicas. É sabido que os expoentes
de Lyapunov podem ser muito sensíveis como funções do cociclo. Exemplo disto é
o resultado de Bochi-Mañé que mostra que todo SL(2,R)-cociclo contínuo que não
é uniformemente hiperbólico pode ser aproximado por outro com expoentes nu-
los. Mostrarei que o conjunto dos SL(2,R)-cociclos “fiber-bunched”; com expoente
de Lyapunov não nulos, sobre um difeomorfismo parcialmente hiperbólico, é um
aberto. Este é um trabalho conjunto com Lucas Backes e Mauricio Poletti.
O segundo tipo de resultados trata de expoentes de Lyapunov de cociclos localmente
constantes associados a distribuções de probabilidade com suporte não compacto
em SL(2,R). Bocker-Viana provaram que, para distribuições com suporte compacto,
os expoentes variam continuamente. Analizarei o comportamente dos expoentes de
Lyapunov quando as medidas têm suporte não compacto, mostrando que neste caso
tem-se semi-continuidade com a topologia de Wasserstein, mas não na topologia
fraca*. Além disso, não há continuidade mesmo na topologia de Wasserstein.
vi
-
Acknowledgement
First and foremost I would like to express my sincere gratitude to my advisor Prof.
Marcelo Viana for the continuous support of my Ph.D study and related research, for
his patience, motivation, and immense knowledge. He supported me with prompt-
ness and care, and has always been patient and encouraging even during tough
times in the Ph.D. pursuit.
Besides my advisor, I would like to thank the rest of my thesis committee: Prof.
Carlos Bocker Neto, Prof. Karina Marín, Prof. Lorenzo Díaz, Prof. Enrique Pujals,
and Prof. Jacob Palis, for their insightful comments and also, for the hard question
which incented me to widen my research from various perspectives. I would also
like to thank Gugu, Emanuel Carneiro and Jorge Vitorio for all the help during my
masters and Ph.D. Furthermore, I am very grateful to William Alvarado this work
would not exist with out your support and encouragement.
The work on this thesis was supported by the Conselho de Desenvolvimento Cien-
tífico e Tecnológico (CNPq) and the Universidad de Costa Rica (UCR). My dearest
thanks for their finantial support.
I have been very privileged to get to know and to collaborate with many other great
people who became friends over the last several years. I thank my UCR collegues,
Alberto Fonseca, Raul Bolaños, Samaria Montenegro, Rafael Zamora, Esteban Se-
gura, Dario Mena,and Iván Ramirez for the stimulating discussions, for the sleep-
less nights we were working together before deadlines, and for all the fun we had
during the years. Special thanks to Alejandra Guerrero, Karen Guevara, Alejandra
Camacho, Jeff Maynard, Jennifer Loria and Oscar Quesada, my costarrican comu-
nity during my time at Rio. Without you I would probably come back home long
time before.
My time at IMPA was made enjoyable in large part due to the many friends and
groups that became a part of my life. My friends Mateus, Diogo, Allan, Gabrielle,
Felipe, Marlon, Daniel, Clarena, Jamerson, Sandoel, Alex, Viviana and, Hudson. To
my latin family Heber, Yulieth, David, Midory, Cani, Betina, Nico, Inocencio, Laura,
Plinio and Vanessa, thank you for your friendship and support.
I have greatly enjoyed the opportunity to work with Mauricio, Ermerson, Cata and
Yaya. Thank you for teaching me so much in our join research and long hours of
vii
-
study. You guys taught me a great deal about dynamics and will be in my heart
forever, no matter wherever we are.
I am very grateful to Gabriela Araya and Miguel Garcia for their friendship over the
years. Thank you for being there despite the distance.
Last but not the least, I would like to thank my family: my husband and my mom.
This work is for you and because of you. I woudln’t be here if it wasn’t for your love
and support when I have needed it the most.
Adriana Sánchez
IMPA
August, 2018
viii
-
Contents
1 Introduction 1
1.1 Structure of the work . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2 Definitions and statements 5
2.1 Linear cocycles and Lyapunov exponents . . . . . . . . . . . . . . . . 5
2.2 Partial hyperbolicity . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.3 Wasserstein topology . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.4 Main results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.4.1 Partially hyperbolic base dynamics . . . . . . . . . . . . . . . 8
2.4.2 Measures with non compact support . . . . . . . . . . . . . . 9
3 Partially hyperbolic base dynamics 11
3.1 Preliminary results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.1.1 Holonomies and disintegrations . . . . . . . . . . . . . . . . . 12
3.1.2 Accessibility and holonomies . . . . . . . . . . . . . . . . . . 13
3.1.3 PSL(2,R) cocycles and invariant measures in P1 . . . . . . . 18
3.2 Proof of Theorem A . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.2.1 Continuity and convergence of conditional measures . . . . . 22
3.2.2 Excluding the atomic case with a bounded number of atoms . 25
3.2.3 Conclusion of the proof . . . . . . . . . . . . . . . . . . . . . 27
3.3 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.3.1 Proof of Theorem B . . . . . . . . . . . . . . . . . . . . . . . . 30
3.3.2 Random product cocycles . . . . . . . . . . . . . . . . . . . . 30
4 Probability distributions with non compact support 33
4.1 Semicontinuity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
4.1.1 Semicontinuity counterexample with weak* topology . . . . . 33
4.1.2 Semicontinuity relative to the Wasserstein topology . . . . . . 34
4.2 Examples of discontinuity . . . . . . . . . . . . . . . . . . . . . . . . 39
4.2.1 Proof of Theorem D . . . . . . . . . . . . . . . . . . . . . . . 39
4.2.2 Discontinuity example in SL(2,R)5 . . . . . . . . . . . . . . . 41
4.2.3 Discontinuity example in GL(2,R)2 . . . . . . . . . . . . . . . 46
Bibliography 49
ix
-
1IntroductionThe theory of linear cocycles goes back to the works of Furstenberg, Kesten [15,
14] and Oseledets [22]. The simplest examples of linear cocycles are given by
derivative transformations of smooth dynamical systems. The cocycle generated by
A(x) = Df(x) over f is called the derivative cocycle. Taking as an example the hy-
perbolic theory of Dynamical Systems where one can understand certain dynamical
properties of f by studying the action ofDf on the tangent space, one can hope that
by studying properties of linear cocycles one can also deduce some properties of f .
Nevertheless, the notion of linear cocycle is much more general and flexible, and
arises naturally in many other situations as in the spectral theory of Schrödinger
operators, for instance.
In the present work we are interested in the asymptotic behavior of An(x). Thus,
we are interested in understanding certain regularity properties of Lyapunov expo-
nents. The Lyapunov exponents are quantities that measure the average exponen-
tial growth of the norm iterates of the cocycle along invariant subspaces on the
fibers. They describe the chaotic behavior of the system. For example, a strictly
positive maximal Lyapunov exponent is synonymous of exponential instability. It is
an indication that the system modeled by the cocycle behaves chaotically, and the
maximal Lyapunov exponent measures the chaos. These objects are one of the most
fundamental notions in dynamical systems.
It is well known that, in general, Lyapunov exponents can be very sensitive as func-
tions of the cocycle. For instance, Bochi [7, 8] proved that in the space of SL(2,R)-
valued continuous cocycles over an aperiodic map, if a cocycle is not hyperbolic,
then it can be approximated by cocycles with zero Lyapunov exponents. In partic-
ular, there are cocycles with positive Lyapunov exponents that are accumulated by
cocycles with zero Lyapunov exponents.
Furthermore, when the base dynamic is far from being hyperbolic, for example,
when f is a rotation on the circle, Wang and You [26], showed that having non-
zero Lyapunov exponents is not an open property even in the C∞ topology.
Bocker and Viana [9] constructed an example over a hyperbolic map showing that
the same phenomenon can happen in the Hölder realm. In order to construct their
example, Bocker and Viana exploited the fact that the cocycle is not fiber-bunched.
In fact, it was shown by Backes, Butler and Brown [5] that in the fiber-bunched set-
ting over a hyperbolic map the Lyapunov exponents vary continuously with respect
to the cocycle and, in particular, cocycles with positive Lyapunov exponents can not
be approximate by cocycles with zero Lyapunov exponents.
1
-
In the first part of this work we are interested in understanding the case when the
cocycle still have some regularity properties, namely, it is fiber-bunched but the base
dynamics exhibit some mixed behaviour of hyperbolicity and non-hyperbolicity, that
is, the map f is partially hyperbolic. More precisely, f over a compact manifold M
is such that there exists a nontrivial, Df -invariant splitting of the tangent bundle
TM = Es ⊕ Ec ⊕ Eu and a Riemannian metric on M such that vectors in Es are
uniformly contracted by Df in this metric, vectors in Eu are uniformly expanded,
and the expansion and contraction rates of vectors in Ec is dominated by the cor-
responding rates in Eu and Es, respectively. Furthermore, we say that f is center
bunched if the contraction rate on Es and the expanding rate on Eu are uniformly
bounded by the product of the contracting and expanding rates on Ec.
An su-path in M is a concatenation of finitely many subpaths, each of which lies
entirely in a single leaf of Ws or Wu. We say that f is accessible if any point in M
can be reached from any other along an su-path.
We show that if f is chaotic enough and A is fiber-bunched then the Bochi phe-
nomenon can not occur. That is, (see Chapter 2 for detailed definitions),
Theorem A. If (f, µ) is a volume preserving partially hyperbolic accessible and center-
bunched diffeomorphism over M , and A : M → SL(2,R) is a Hölder continuous fiber-
bunched map with nonvanishing Lyapunov exponents. Then A can not be accumulated
by cocycles with zero Lyapunov exponents.
Moreover, we show that the accessibility assumption in the previous result is neces-
sary.
Theorem B. There exists a volume preserving partially hyperbolic and center-bunched
diffeomorphism f and a Hölder continuous fiber-bunched map A with non-zero Lya-
punov exponents which is approximated by cocycles with zero Lyapunov exponents.
Notice that in the results mention above we consider the Lyapunov exponents as
funtions of the cocycle. The purpose of the second part of this work is to study the
continuity and semicontinuity of the Lyapunov exponents respect to measures of
non-compact support. That is, for the second part of this work we will consider the
Lyapunov exponents are functions of the measures. Moreover, we restrict our study
to the particular case of products of random matrices i,e, when the linear cocycle is
given by A((αk)k) = α0 where αk ∈ SL(2,R) for every k ∈ Z.
Our main result reads as follows (see Theorem 2.4.3 and Theorem 2.4.4 for a pre-
cise statement):
Theorem C. The function p 7→ λ+(p) is upper semi-continuous relative to the Wasser-
stein topology but not with the weak* topology. The same remains valid for p 7→ λ−(p)
with lower semi-continuity.
2 Chapter 1 Introduction
-
The main problem regarding the weak* topology is that it is defined in terms of
bounded continuous or bounded Lipschitz functions. That is, we have that µn∗
−→ µ
if∣
∣
∣
∣
∫
ψdµn −∫
ψµ
∣
∣
∣
∣
→ 0,
for all bounded continuous (or Lipschitz) functions (see [21, Chapter 2]). In the
case of probability measures with non compact support we don’t have a bounded co-
cycle so we can not guarantee the semicontinuity as in the compact case. However,
convergence in the Wasserstein topology, as we will see in Section 2.3, is equivalent
to the convergence of integrals of Lipschitz, not necessarily bounded, functions.
Moreover, this topology is defined over borel measures with finite first moment and
the convergence implies convergence of the moments of order 1. These two prop-
erties are the ones that allow us to control the convergence outside a compact set
and prove the semicontinuity.
Regarding continuity of Lyapunov exponents we prove the following (see Theorem
2.4.5 for a precise statement):
Theorem D. The function p 7→ λ+(p) is not continuous in the Wasserstein Topology.
The same remains valid for p 7→ λ−(p).
For the proof of this theorem we present an example of measures with vanishing
Lyapunov exponent converging to one with strictly positive exponent. The main
idea is to take a measure with a countable support containing only hyperbolic ma-
trices and create a sequence by replacing one element of the original support by a
rotation that exchanges vertical and horizontal axes. However, the support of the
measures constructed this way move further apart from the original support.
The same result can be obtained by considering more variables in the alphabet,
for example SL(2,R)5. Hence, instead of one (big) rotation we can decompose it
into several small rotations in each coordinate allowing us to leave the supports
arbitrarily close.
1.1 Structure of the work
The present work is divided in three parts:
• In Chapter 2 we present basic definitions and some preliminary results in
order to give a precise statement of Theorems A through D.
• Chapter 3 is devoted to the proof of the results regarding partially hyperbolic
diffeomorphims. Section 3.1 has the preliminary results we are going to need
in the proof. The second section, Section 3.2, presents the proof of Theorem
A while the proof of Theorem B is presented in Section 3.3.
1.1 Structure of the work 3
-
• In Chapter 4 we focus on the study of probability measures with non compact
support. The first section is devoted to the study of semicontinuity while the
second one focuses on analyze the continuity.
4 Chapter 1 Introduction
-
2Definitions and statements
2.1 Linear cocycles and Lyapunov exponents
Let (M,B, µ) be a measurable space and f an invertible measure preserving trans-
formation f : (M,µ) → (M,µ). A measurable function A : M → GL(2,R) gives the
dynamically defined products
An(x) =
A(fn−1(x)) . . . A(f(x))A(x), if n > 0,
Id, if n = 0,
(A−n(fn(x)))−1 = A(fn(x))−1 . . . A(f−1(x))−1, if n < 0.
(2.1)
The linear cocycle defined by A over f is the transformation
F : M × R2 → M × Rd (x, v) 7→ (f(x), A(x)v),
where its n-th iterate is given by Fn(x, v) = (fn(x), An(x)v).
Let L1(µ) denote the space of µ-integrable functions onM and suppose that log+ ‖A±1‖
belongs to L1(µ). It follows from the sub-additive ergodic theorem of Kingman [19],
that the limits
λ+(A,x) = limn→∞
1
nlog ‖An(x)‖,
λ−(A,x) = limn→∞
1
nlog ‖A−n(x)‖−1,
exist for µ-almost every x ∈ M . We call such limits Lyapunov exponents. Moreover,
when λ+(A,x) > λ−(A,x) it follows from the well-known theorem of Oseledets
[22] that there exists a decomposition R2 = Eu,Ax ⊕ Es,Ax into vector subspaces
depending measurably on x such that for µ-almost every point
A(x)Eu,Ax = Eu,Af(x),
λ+(A,x) = limn→∞
1
nlog ‖An(x)v‖,
for every non-zero v ∈ Eu,Ax . Equivalently for s and λ−(A,x). This decomposition
is called the Oseledets decomposition. Furthermore, the Lyapunov exponents are f -
invariant, so if µ is ergodic it implies that they are constant for µ-almost every point
x. In this case we write λ+(A,x) = λ+(A,µ) and λ−(A,x) = λ−(A,µ).
5
-
2.2 Partial hyperbolicity
Let f : M → M be a Cr, r ≥ 2, diffeomorphism defined on a compact manifold M ,
f is said to be partially hyperbolic if:
1. There exists a non-trivial splitting of the tangent bundle TM = Es ⊕Ec ⊕Eu
invariant under the derivative Df ;
2. There exist a Riemannian metric ‖ · ‖ on M , such that we have positive con-
tinuous functions ν, ν̂, γ, γ̂ with ν, ν̂ < 1 and ν < γ < γ̂−1 < ν̂−1 such that,
for any unit vector v ∈ TxM ,
‖Df(x)v‖ < ν(x) if v ∈ Es(x),
γ(x)
-
for all x, y ∈ M where ‖A‖ denotes the operator norm of a matrix A, that is,
‖A‖ = sup {‖Av‖/‖v‖; ‖v‖ 6= 0} .
Let Hα(M) denote the space of all such α-Hölder continuous maps. We endow this
space with the α-Hölder topology which is generated by the norm
‖A‖α = supx∈M
‖A(x)‖ + supx 6=y
‖A(x) −A(y)‖
dist(x, y)α.
We say that the linear cocycle generated by A over f is fiber-bunched if
‖A(x)‖‖A(x)−1‖ν(x)α < 1 and ‖A(x)‖‖A(x)−1‖ν̂(x)α < 1
for every x ∈ M . Since our base dynamics f is going to be fixed, we simply say that
A is fiber-bunched. Observe that this is an open condition in Hα(M).
2.3 Wasserstein topology
Let (M,µ) and (N, ν) be two probability spaces. Coupling µ and ν means construct-
ing a measure π on M ×N , such that π projects to µ and ν on the first and second
coordinate respectively. When µ = ν we call π a self-coupling.
If (M,d) is a Polish metric space, for any two probability measures µ, ν on M , the
Wasserstein distance between µ and ν is defined by the formula
W1(µ, ν) = infπ∈Π(µ,ν)
∫
Md(x, y)dπ(x, y), (2.2)
where the infimum is taken over the set Π(µ, ν) which denotes the set of all the
couplings of µ and ν.
The Wasserstein space is the space of probability measures which have a finite
moment of order 1. By this we mean the space
P1(M) :=
{
µ ∈ P (M) :∫
Md(x0, x)dµ(x) < +∞
}
,
where x0 ∈ M is arbitrary and P (M) denotes the space of Borel probability mea-
sures on M . This does not depend on the choice of the point x0, and W1 defines a
finite distance on it (see [1, Chapter 7]).
An important property of the Wasserstein topology is the Kantorovich duality. It
establishes that
W1(µ, ν) = sup
{∫
Mψdµ −
∫
Mψdν
}
,
where the supremum on the right is over all 1-Lipschitz functions ψ.
2.3 Wasserstein topology 7
-
The next definition characterizes the convergence in the Wasserstein space P1(M).
From now on the notation µkW
−→ µ means that µk converges in the Wasserstein
topology, while µk∗
−→ µ means that µk converges in the weak* topology.
Definition 2.3.1. [25, Definition 6.8] Let (M,d) be a Polish metric space. Let (µk)k∈Nbe a sequence of probability measures in P1(M) and let µ be another element of P1(M).
Then µk is said to converge in the Wasserstein topology to µ, if one of the following
equivalent properties is satisfied for some (and then any) x0 ∈ M :
1. µk∗
−→ µ and∫
d(x0, x)dµk(x) →∫
d(x0, x)dµ(x);
2. µk∗
−→ µ and
lim supk→∞
∫
d(x0, x)dµk(x) ≤∫
d(x0, x)dµ(x);
3. µk∗
−→ µ and
limR→∞
lim supk→∞
∫
d(x0,x)≥Rd(x0, x)dµk(x) = 0;
4. For all continuous functions ϕ with |ϕ(x)| ≤ C(1 + d(x0, x)), C ∈ R, one has
∫
ϕ(x)dµk(x) →∫
ϕ(x)dµ(x).
A crucial fact is that the Wasserstein distance W1 metrizes the convergence in the
Wasserstein topology in P1(M). In other words, µkW
−→ µ if and only if W1(µk, µ) →
0. This equivalence also implies that W1 is continuous on P1(M) (see [25, Theorem
6.18]).
Theorem 2.3.1 (Topology in P1(M)). Let (M,d) be a Polish metric space. Then the
Wasserstein distance W1, metrizes the convergence in the Wasserstein topology in the
space P1(M). Moreover, with this metric P1(M) is also a complete separable metric
space and, any probability measure can be approximated by a sequence of probability
measures with finite support.
2.4 Main results
With the definitions mention in the last sections we now proceed to properly state
the results mention in Chapter 1.
2.4.1 Partially hyperbolic base dynamics
Let f : M → M be a partially hyperbolic diffeomorphism on a compact manifold M
and µ a probability measure in the Lebesgue class of M . By Lebesgue class we mean
8 Chapter 2 Definitions and statements
-
the set of measures that are generated by a volume form. Moreover, we say that f
is volume preserving if it preserves some probability measure in the Lebesgue class
of M.
The main results regarding partially hyperbolic diffeomorphisms are the follow-
ing.
Theorem 2.4.1. Let f : M → M be a Cr, r ≥ 2, partially hyperbolic, volume pre-
serving, center bunched and accessible diffeomorphism defined on a compact manifold
M . Also, let µ be an ergodic f -invariant measure in the Lebesgue class. If A ∈ Hα(M)
is fiber-bunched and λu(A,µ) > λs(A,µ) then A can not be accumulated by cocycles
with zero Lyapunov exponents.
We observe that a similar result can be stated in terms of GL(2,R)-valued cocy-
cles changing ‘cocycles with zero Lyapunov exponents’ by ‘cocycles with just one
Lyapunov exponent’.
Indeed, the accessibility property guarantees connectedness of M . Hence, by con-
tinuity of A either, det(A(x)) > 0 for every x ∈ M or det(A(x)) < 0 for every
x ∈ M . Suppose we are in the first case (the other case can be easily deduced
from this one). Then, given A : M → GL(2,R) consider gA : M → R the map de-
fined by gA(x) = (detA(x))1
2 and B : M → SL(2,R) such that A(x) = gA(x)B(x).
Therefore,
λu/s(A,µ) = λu/s(B,µ) +
∫
log(gA(x)) dµ(x),
and consequently,
λu(A,µ) = λs(A,µ) ⇐⇒ λu(B,µ) = 0 = λs(B,µ).
We also present an example showing that the accessibility assumption in the previ-
ous theorem is necessary. More precisely,
Theorem 2.4.2. There exists a volume preserving partially hyperbolic and center-
bunched diffeomorphism f and a Hölder continuous fiber-bunched map A with non-
zero Lyapunov exponents which is approximated by cocycles with zero Lyapunov expo-
nents.
2.4.2 Measures with non compact support
Let M = SL(2,R)Z and, let f : M → M be the shift map over M defined by
(αn)n 7→ (αn+1)n.
Consider the function
A : M → SL(2,R), (αn)n 7→ α0,
2.4 Main results 9
-
and we define its n-th iterate, the product of random matrices, by
An((αk)k) = αn−1 · · ·α0.
Given an invariant measure p in SL(2,R) we can define µ = pZ which is an invariant
measure in M .
It is a well-known fact that when the measures have compact support, the Lyapunov
exponents are semicontinuous with the weak* topology (see for example [24, Chap-
ter 9]). However, in the non compact setting this is no longer true. If they were
semicontinuos then every measure with vanishing Lyapunov exponents would be a
point of continuity. The next theorem shows that this is not the case.
Theorem 2.4.3. There exist a measure p and a sequence of measures (qn)n on SL(2,R)
converging to q in the weak* topology, such that λ+(qn) ≥ 1 for n large enough but
λ+(q) = 0.
Consider in SL(2,R) the metric given by
d(α, β) = ‖α− β‖ + ‖α−1 − β−1‖.
Since the space SL(2,R) is a Polish metric space with this metric we can consider
the Wasserstein topology in P1(SL(2,R)).
The Wasserstein topology is stronger than the weak* topology, as mentioned in
Definition 2.3.1. The principal consequence of the convergence in the Wasserstein
topology is that it implies convergence of the moments of order 1. This allow us to
control the weight of integrals outside compact sets and, proof semi-continuity of
the Lyapunov exponents in P1(SL(2,R)). We are thus led to the following result.
Theorem 2.4.4. The function defined on P1(SL(2,R)) by p → λ+(p) is upper semi-
continuous relative to the Wasserstein topology. The same remains valid for the func-
tion p → λ−(p) with lower semi-continuity.
Finally, we will present a construction of discontinuity points of the Lyapunov expo-
nents as functions of the probability measure, relative to the Wasserstein topology.
This implies that the Wasserstein topology is not enough to guarantee continuity of
the Lyapunov exponents.
Theorem 2.4.5. There exist a measure q and a sequence of measures (qn)n on SL(2,R)
converging to q in the Wasserstein topology, such that λ+(qn) = 0 for all n ∈ N but
λ+(q) > 0.
10 Chapter 2 Definitions and statements
-
3Partially hyperbolic basedynamics
The aim of this chapter is to give a proof for Theorem A and Theorem B. Let us
give an outline of the proof of Theorem A, it goes by contradiction. Assume there
exist a sequence {Ak}k with λu(Ak, µ) = λs(Ak, µ) = 0 for every k ∈ N and such
that Ak converges to A in the Hölder topology. The basic strategy is to consider
the projective cocycles FAk , FA : M × P1 → M × P1 defined by (A, f) and (Ak, f)
respectively, and to analyze the probability measures m and mk on M ×P1 that are
invariant under FA and FAk respectively, and project down to µ on M .
The accessibility condition allow us to define, in Section 3.1.2, holonomy maps HAγfor every su-path γ on every point x ∈ M . Using a result of Avila, Santamaria,
Viana in [3] we show in Section 3.2.1 that the Oseledets decomposition is invariant
by those holonomy maps. Thus, we can separate our proof in two cases:
In the first case we assume there exist x ∈ M and a non trivial su-loop γ on x
such that HAγ is hyperbolic. Hence, for k sufficiently large the holonomy maps HAkγ
are also hyperbolic and the conditional measures mkx have at most two atoms. In
Section 3.2.1 we prove uniform convergence of the conditional measures and, in
Section 3.2.2 we use it to prove that mkx can not have a finite number of atoms
contradicting this case.
In the second case we assume that for every point x and every su-loop γ at x that
HAγ = id. Section 3.1.3 gives a characterization of the sequence of holonomy maps
for the cocycles Ak. This characterization give us again two cases: either there exist
a nontrivial su-loop γ at some point x such that HAkγ is hyperbolic for infinitely
many k or they are the identity for every su-loop for some k large enough. The
hyperbolic case is solved as before, for the identity case, we can perform a change
of coordinates that makes the cocycles A and Ak constant without changing its
Lyapunov exponents, this is explained in Section 3.1.2. The proof concludes by
establishing the convergence of the Lyapunov exponents for the constant cocycles.
This is done in Section 3.2.3.
Section 3.3 contains the proof of Theorem B as well as another construction show-
ing that we have a fiber-bunched cocycle A over a partially hyperbolic and center-
bunched map f with arbitrarily large Lyapunov exponent λ+ which can be approxi-
mated by cocycles with zero Lyapunov exponents.
The results presented in this chapter are the product of a joint work with Lucas
Backes and Mauricio Poletti in [6].
11
-
3.1 Preliminary results
In this section we recall some classical notions and present some useful results that
are going to be used in the proof of our main theorem. Along this chapter we
consider f : M → M , A ∈ Hα(M) and µ be as in Theorem A.
3.1.1 Holonomies and disintegrations
Given x, y ∈ M , we define the equivalence relation x ∼s y by y ∈ Ws(x). Observe
that this is f -invariant, that is, if x ∼s y then f(x) ∼s f(y). Given L > 0, we write
x ∼sL y when there exist a sequence of points x = z0, ..., zn = y such that zi ∼s zi+1
and dWs(zi+1, zi) ≤ L for every i = 1, . . . , n− 1.
For every pair of points x, y ∈ M satisfying that x ∼s y, the fiber-bunched assump-
tion assures that the limit
Hs,Axy = limn→+∞An(y)−1 ◦ An(x)
exists (see [3, Proposition 3.2]). Moreover, by [3, Remark 3.4] we know that for
every L > 0, the map
(x, y,A) → Hs,Axy
is continuous on WsL ×Hα(M), where WsL = {(x, y) ∈ M ×M ;x ∼
sL y}.
The family of maps Hs,Axy is called an stable holonomy for the cocycle (A, f). Also by
[3, Proposition 3.2] we have for x ∼sL y and z ∼sL y,
• Hs,Axx = id,
• Hs,Axy = Hs,Azy ◦H
s,Axz ,
• ‖HAxy‖ ≤ N for some N > 0 and,
• Hs,Afj(x)fj (y)
= Aj(y)Hs,Axy Aj(x)−1 ∀j ≥ 0.
Dually, we write x ∼u y if y ∈ Wu(x) and, we define the unstable holonomy Hu,Axyas the stable holonomies for (A−1, f−1). That is
Hu,Axy = limn→−∞An(y)−1 ◦ An(x)
whenever x and y are on the same strong-unstable leaf. Even more, (x, y,A) →
Hu,Axy is continuous on WuL ×H
α(M), where WuL = {(x, y) ∈ M ×M ;x ∼uL y}.
We say that a measure m on M × P1 projects on µ if π∗m = µ where π is the
canonical projection π : M × P1 → M . By the Ergodic Decomposition theorem
[21, Theorem 5.1.3] any such measure admits a disintegration with respect to the
partition {{x}×P1}x∈M and the measure µ, that is, there exists a family of measures
{mx}x∈M on {{x} × P1}x∈M so that for every measurable B ⊂ M × P1,
12 Chapter 3 Partially hyperbolic base dynamics
-
• The map x → mx(B) is measurable,
• mx({x} × P1) = 1 and,
• m(B) =∫
M mx(B ∩ ({x} × P1))dµ(x).
Moreover, such disintegrantion is essentially unique [23]. Identifying each fiber
{x} × P1 with P1, we can think of x → mx as a map from M to the space of
probability measures on P1 endowed with the weak* topology.
Given B ∈ GL(2,R) we write PB : P1 → P1 for the induced projective map. That
is, the map defined by PB[v] = [Bv].
The projective cocycle associated to A is the map FA : M × P1 → M × P1 given by
FA(x, v) = (f(x),PA(x)[v]).
Observe that m is FA-invariant if and only if (PA(x))∗mx = mf(x) for µ-almost
every point x ∈ M .
We say that a FA-invariant measure m projecting on µ is essentially s-invariant if
there exists a total measure set M s ⊂ M such that for every x, y ∈ M s satisfying
x ∼s y we have(
Hs,Axy
)
∗mx = my.
Such measure m is also known as an s-state. One speaks of s-invariant if M s = M .
Analogously, we say that m is essentially u-invariant (or an u-state) if the same is
true replacing stable by unstable in the previous definition. We say that m is es-
sentially su-invariant if it is simultaneously essentially s-invariant and essentially
u-invariant. The main property of essentially su-invariant measures is the follow-
ing
Proposition 3.1.1. [3, Theorem B] If λ+(A,µ) = λ−(A,µ), any FA-invariant mea-
sure m projecting on µ admits a disintegration {mx}x∈M for which M s = Mu = M .
Moreover, accessibility of f implies that mx depends continuously on the base point
x ∈ M in the weak* topology.
3.1.2 Accessibility and holonomies
An su-path from x to y is a path connecting x and y which is a concatenation of
finitely many subpaths, each of which lies entirely in a single leaf of Ws or a single
leaf of Wu. Every sequence of points x = z0, z1, . . . , zn = y, with the property that ziand zi+1 lie in the same leaf of either W s or W u , for i = 0, ..., n−1 defines a unique
su-path (see Figure 3.1). If in addition x = y, then the sequence determines an su-
loop or a closed su-path (see Figure 3.2). With these terminology the accessibility of
f means that any point in M can be reached from any other along an su-path.
3.1 Preliminary results 13
-
Wu
Ws
bxb
z1
bz2 b z3
b
z4b z5
b y
Figure 3.1: su-path
We define the concatenation of two su-paths γ1 given by z0, . . . , zn and γ2 given by
z′0, z′1, . . . , z
′m, with z
′0 = zn, by γ1 ∧ γ2 which is the su-path given by the sequence
of points z0, . . . , zn, z′1, . . . , z′m.
We say that an su-path γ defined by the sequence x = z0, z1, . . . , zn = y is a (K,L)-
path if n ≤ K and dW∗(zi+1, zi) ≤ L for every i = 1, . . . , n − 1 where dW∗ is
the distance induced by the Riemannian structure on the submanifold W∗ for ∗ =
s, u. Observe that, by the compactness of M and continuity of stable (unstable)
manifolds of bounded size, the space of (K,L)-paths is compact. In particular,
Wilkinson [27] proved that accessibility implies uniform accessibility:
Lemma 3.1.2. [27, Lemma 4.5] There exist constants K and L such that every pair
of points in M can be connected by an (K,L)-path.
Consider K and L given by the lemma above. If γ is the su-path defined by the
sequence z0, z1, . . . , zn then we write HAγ = H∗,Azn−1zn ◦ . . . ◦H
∗,Az0z1 for ∗ ∈ {s, u}.
Let us assume that HAγ = id for every (3K,L)-loop. This implies that the same
remains valid for every su-loop. Indeed, observe initially that any su-loop γ can be
transformed into an su-loop with legs of size at most L just by breaking one “large”
leg into several with smaller sizes. Thus, it is enough to consider su-loops with legs
of size at most L.
If γ is a (2K,L)-path from x to y then, by Lemma 3.1.2, there exists a (K,L)-path
γ′ from x to y. If −γ′ denotes the path γ′ with opposite orientation then γ ∧ (−γ′)
is a (3K,L)-loop and
HAγ ◦ (HAγ′)
−1 = HAγ ◦HA−γ′ = H
Aγ∧(−γ′) = id .
14 Chapter 3 Partially hyperbolic base dynamics
-
Wu
Ws
bxb
z1
bz2 b z3
b
z4b z5
b z6
b
z9
b z8
b
z7
Figure 3.2: su-loop
Hence, HAγ = HAγ′ .
Now, taking any su-loop γ with an arbitrary number of legs whose lengths are at
most L we can decompose it as γ = γ1 ∧ · · · ∧γk, where every γi is a (K,L)-path. In
particular, γk−1 ∧ γk is a (2K,L)-path and by the previous argument we can replace
it by a (K,L)-path γ′k−1 with the same starting and ending points and, so that
HAγk−1∧γk = HAγ′
k−1.
Thus, taking γ′ = γ1 ∧· · ·∧γk−2 ∧γ′k−1 we have that γ and γ′ have the same starting
and ending points and HAγ = HAγ′ . Repeating this procedure a finite number of
times we get some (K,L)-loop γ′′ such that HAγ = HAγ′′ = id. Finally, we conclude
that HAγ = id for every su-loop proving our claim. An example of this process is
shown in Figure 3.3 where L = 4 and K = 2.
As a consequence we get that if γ is an su-path connecting x and y then HAγ does not
depend on γ. In fact, if γ1 and γ2 are su-paths connecting x and y then γ1 ∧ (−γ2) is
an su-loop and thus
HAγ1 ◦ (HAγ2)
−1 = HAγ1 ◦HA−γ2 = H
Aγ1∧(−γ2) = id .
Let us denote this common value simply byHAxy. From the properties of the holonomies
and the fact that any two points x, y ∈ M can be connected by a (K,L)-path it fol-
lows that
• HAyzHAxy = H
Axz,
• A(y)HAxy = HAf(x)f(y)A(x),
3.1 Preliminary results 15
-
Wu
Ws
bx = z0
bz1bz2
bz3 b
z4
bz5
b
z7
b
z8
b
z9
b
z10
b
z11
b z6
b
z15
b z14
b
z13
γ
b
z12
(a)Original su-loop γ
Wu
Ws
bx = z0
bz1bz2
bz3 b
z4
bz5
b
z8
b
z9
b
z10
b
z11
b
z12
b z7
bb
b
bz6
b
z13
γk
bγ′
k−1
γk−1
(b)Construction of γ′k−1
Wu
Ws
b
x = z0
bz1bz2
bz3 bz4
bz5
b
z8
b
z9
b
z10
b
z11
b
z12
b z7
b
z16
b
z15
b
z14
b z6
b
z13
b
z′5
γ′′
(c)Final su-loop γ′′
Figure 3.3: Process to transform an arbitrary su-loop into a (3K,L)-loop.
16 Chapter 3 Partially hyperbolic base dynamics
-
• A → HAxy is uniformly continuous for any pair of points x, y ∈ M and
• ‖HAxy‖ ≤ N for some N > 0 and any x, y ∈ M .
Fix x ∈ M and, given y ∈ M , consider the following transformation
Â(y) = HAf(y)xA(y)HAxy. (3.1)
We are going to prove that this change of coordinates makes the cocycle (A, f)
constant without changing its Lyapunov exponents.
Notice that by the properties of the holonomies
Â2(y) = Â(f(y))Â(y) = HAf2(y)xA(f(y))HAxf(y)H
Af(y)xA(y)H
Axy,
and consequently Â2(y) = HAf2(y)xA2(y)HAxy. More generally, Â
n(y) = HAfn(y)xAn(y)HAxy
for every n ∈ N. Hence, (Â, f) and (A, f) have the same Lyapunov exponents. In-
deed, notice that
λ+(Â, µ) = limn→∞
1
nlog ‖Ân(y)‖
= limn→∞
1
nlog ‖HAfn(y)xA
n(y)HAxy‖
≤ limn
1
n(2 logN + log ‖An(y)‖)
= λ+(A,µ).
Similarly, since An(y) = HAxfn(y)Ân(y)HAyx we have that λ+(A,µ) ≤ λ+(Â, µ). Thus,
we proved our claim.
Moreover, for any z, y ∈ M ,
Â(z)−1Â(y) =(
HAf(z)xA(z)HAxz
)−1HAf(y)xA(y)H
Axy
= HAzxA(z)−1HAxf(z)H
Af(y)xA(y)H
Axy
= HAzxA(z)−1HAf(y)f(z)A(y)H
Axy
= HAzxA(z)−1A(z)HAyzH
Axy
= HAzxHAyzH
Axy
= HAzxHAxz
= id .
In particular, Â is constant and consequently its largest Lyapunov exponent is the
logarithm of the norm of the greatest eigenvalue of Â. Summarizing, if HAγ = id for
every (3K,L)-loop γ then we can perform a change of coordinates that makes the
cocycle (A, f) constant without changing its Lyapunov exponents.
3.1 Preliminary results 17
-
3.1.3 PSL(2,R) cocycles and invariant measures in P1
Let PGL(2,R) be the projective linear group, that is the induced action of the
general linear group on the associated projective space P1. Let π : R2 \{0} → P1 be
the natural projection given by v 7→ [v]. Each automorphism P̃ ∈ GL(2,R) induces
a projective transformation P ∈ PGL(2,R) through P [v] = [P̃ v].
On the other hand, endomorphisms of R2 (i.e. linear maps) do not project, in
general, to self maps of P1. Nevertheless, it was pointed out by Furstenberg [15]
that the space of projective maps has a natural compactification. If Q̃ ∈ End(R2)
is a linear transformation of rank r > 0 with kernel ker Q̃ and image Im Q̃ then,
Q̃ determines a quasi-projective transformation Q of P1 given by Q([v]) = [Q̃(v1)]
where v1 is any vector such that v − v1 ∈ ker(Q̃). Observe that Q is defined and
continuous on the complement of the projective subspace kerQ = {[v] : v ∈ ker Q̃},
and its image is ImQ = {[v] : v ∈ Im Q̃}. The number r is called the rank of this
quasiprojective transformation. Rank 1 quasi-projective transformations are quasi-
constant maps, each of them is undefined on a hyperplane in P1 and its image is a
single point. We refer the reader to [17], [13] or [10] for a deeper discussion of
this topic.
The space of quasi-projective transformations inherits a topology from the space
of linear maps, through the natural projection π̃ : Q 7→ Q̃. Clearly, every quasi-
projective transformation Q is induced by some linear map Q̃ such that ‖Q̃‖ = 1. It
follow that the space of quasi-projective transformations is compact for this topol-
ogy (see [13, Theorem 2.83]). In particular, every sequence of projective trans-
formations has a subsequence converging to some quasi-projective transformation
Q.
The next result will be needed in the proof of the proposition below. Its proof can
be found in [10, Lemma 6.1] for projective transformations in CP1. The notion
of quasi-projective maps has been extended to transformations on Grassmannian
manifolds by Gol’dsheid and Margulis [16] and a proof of the following lemma in
this context can be found in [2, Lemma 2.4].
Lemma 3.1.3. If (Pn)n is a sequence of projective transformations converging to a
quasi-projective transformation Q, and (νn)n is a sequence of probability measures
in P1 weakly converging to some probability ν0 with ν0(kerQ) = 0 then (Pn)∗νnconverges weakly to Q∗ν0.
The following result plays an important part in our proof of Theorem A below.
Proposition 3.1.4. Let (Ln)n be a sequence of SL(2,R) matrices converging to id
and, for each n ∈ N let ηn be an Ln-invariant measure on P1 converging in the weak*
topology to 12(δp + δq) for some p, q ∈ P1 with p 6= q. Then for every n sufficiently large
either Ln is hyperbolic or Ln = id.
18 Chapter 3 Partially hyperbolic base dynamics
-
Proof. The proof is by contradiction. We start observing that as Ln converges to the
identity and the trace map is continuous, all the matrices have positive trace for n
sufficiently large. Consequently, if Ln is not the identity we have three possibilities:
the trace tr(Ln) > 2 which means that the matrix Ln is hyperbolic, tr(Ln) < 2 so
the matrix Ln is elliptic or tr(Ln) = 2 and the matrix Ln is parabolic and is non
diagonalizable with both eigenvalues equal to 1.
Elliptic case: Suppose initially that all the matrices Ln have tr(Ln) < 2. Hence
it is conjugated to a rotation of angle θn = arccos(
tr(Ln)2
)
. In particular, for each
n ∈ N there exists Pn ∈ SL(2,R) so that Ln = P−1n RθnPn where Rθn stands for the
rotation of angle θn. Moreover, since tr(Ln)n→+∞−−−−−→ 2, we get that θn
n→+∞−−−−−→ 0.
Now, for each n ∈ N let us consider νn = Pn∗ηn which is an Rθn-invariant measure.
In fact, for every measurable set A ⊂ P1
νn(R−1θnA) = ηn((P
−1n ◦Rθn)A)
= ηn((Rθn ◦ Pn)−1A)
= ηn((Pn ◦ Ln)−1A)
= ηn((L−1n ◦ P
−1n )A)
= ηn(P−1n A)
= νn(A).
We start observing that there exists a subsequence {nj}j so that νnj converges to
Leb where Leb stands for the Lebesgue measure on P1. Indeed, if θn is an irrational
number then we know that the only Rθn-invariant measure is Leb. In particular,
νn = Leb. Thus, if there are infinitely many values of n for which θn is an irrational
number we are done.
Suppose then that θn is a rational number for every n ∈ N. In particular, Rθn is
periodic and denoting by qn its period, since θnn→+∞−−−−−→ 0, we have that qn
n→+∞−−−−−→
+∞.
In what follows we make an abuse of notation thinking of P1 as [0, 1] identifying
the extremes of the interval.
Let ϕ : P1 → R be a continuous map and ε > 0. Since P1 is compact, there exists
δ > 0 so that | ϕ(x) − ϕ(y) |< ε whenever d(x, y) < δ. Thus, taking n ≫ 0 so
that qn > 1δ we get that | ϕ(x) − ϕ(
jqn
)
|< ε for every x ∈ Ij =[
jqn, j+1qn
)
and
j = 0, 1, . . . , qn − 1. In particular,
∣
∣
∣
∣
∣
∣
1
νn([
jqn, j+1qn
))
∫j+1
qn
jqn
ϕdνn − ϕ
(
j
qn
)
∣
∣
∣
∣
∣
∣
< ε. (3.2)
3.1 Preliminary results 19
-
Since νn is Rθn-invariant and
I =qn−1⋃
j=0
Ij =qn−1⋃
j=0
R−jθn (I0),
then νn([
jqn, j+1qn
))
= 1qn for every j = 0, 1, . . . , qn − 1. Summing the expression in
3.2 for j from 0 up to qn − 1 and dividing both sides by qn we get that
∣
∣
∣
∣
∣
∣
∫ 1
0ϕdνn −
1
qn
qn−1∑
j=0
ϕ
(
j
qn
)
∣
∣
∣
∣
∣
∣
< ε.
On the other hand, since ϕ is Riemann integrable,
limn→∞
1
qn
qn−1∑
j=0
ϕ
(
j
qn
)
=
∫
ϕdLeb
which implies that νnn→+∞−−−−−→ Leb as claimed. So, restricting to a subsequence, if
necessary, we may assume that νnn→+∞−−−−−→ Leb.
We now analyze the accumulation points of ηn = P−1n ∗νn. If {P−1n }n stay in a
compact set of SL(2,R) then, taking a subsequence if necessary, we may assume
that there exists P ∈ SL(2,R) so that P−1n → P . In particular, limn→∞ ηn =
P∗Leb which contradicts our assumption since P∗Leb is non-atomic. If∥
∥P−1n∥
∥ → ∞
then we can work on the compactification of quasi-projective transformations. In
particular, restricting to a subsequence, if necessary, we may suppose that Qn =
P−1n /‖P−1n ‖ converges to some quasi-projective map Q. Note that, since ‖P
−1n ‖ =
‖Pn‖ → ∞, ‖Qn‖ = 1 and | detQn| = limn 1/‖Pn‖2 = 0, then Q has rank 1 i.e. is
a quasi-constant map and its image is a single point z (see for instant the proof of
[24, Lemma 6.4]).
Thus, as the kernel has zero Lebesgue measure we can apply Lemma 3.1.3 to con-
clude that
limn→∞
P−1n ∗νn = limn→∞Qn∗νn = Q∗Leb = δz
which is a contradiction. Consequently, Ln may be elliptic only for finitely many
values of n.
Parabolic case: To conclude the proof it remains to rule out the cases when tr(Ln) =
2 and the matrix are non diagonalizable for infinitely many values of n. So, suppose
Ln is non diagonalizable and both of its eigenvalues are 1 for every n. Then by the
Jordan’s normal decomposition we have
Ln = P−1n
(
1 1
0 1
)
Pn
20 Chapter 3 Partially hyperbolic base dynamics
-
for some Pn ∈ GL(2,R). Consequently, the only invariant measure for Ln is atomic
and have only one atom, contradicting the fact that ηnn→+∞−−−−−→ 12(δp + δq). Thus, Ln
can be parabolic and different from id only for finitely many values of n concluding
the proof of the proposition.
Let us consider the projective special linear group given by PSL(2,R) = SL(2,R)/{±Id}.
That is, given A,B ∈ SL(2,R) let us define equivalence relation ∼ given by A ∼ B
if and only if A = B or A = −B. Given A ∈ SL(2,R), let
[[A]] = {B ∈ SL(2,R);B ∼ A}
be the equivalence class of A with respect to ∼. Then, we have that PSL(2,R) =
{[[A]];A ∈ SL(2,R)}. Observe that the norm ‖·‖ on SL(2,R) naturally induces
a norm, which we are going to denote by the same symbol, on PSL(2,R): given
A ∈ SL(2,R),
‖[[A]]‖ := ‖A‖ = ‖−A‖.
Given A : M → SL(2,R) let us consider à : M → PSL(2,R) defined by Ã(x) =
[[A(x)]]. By Kingman’s subadditive ergodic theorem [19] and the ergodicity of µ it
follows that the limit
L(Ã, µ) = limn→+∞
1
nlog ‖Ãn(x)‖
exists and is constant for µ-almost every x ∈ M . In particular, since Ãn(x) =
[[An(x)]] then ‖An(x)‖ = ‖Ãn(x)‖ for every x ∈ M and n ∈ N. Hence, we get
that λu(A,µ) = L(Ã, µ). Another simple observation is that for every v ∈ P1,
[A(x)v] = [Ã(x)v] and, consequently, the action induced by A on P1 coincide with
the action of à on P1.
Moreover, HÃγ = [[HAγ ]] ∈ PSL(2,R) is well defined and have similar properties
with respect to à as those of HAγ with respect to A described in Section 3.1.2.
In particular, a similar conclusion to that of Section 3.1.2 holds for à whenever
HÃγ = [[id]] for every (3K,L)-loop γ: we can perform a change of coordinates
that makes the cocycle (Ã, f) constant without changing L(Ã, µ). Consequently,
denoting this new cocycle by ˆ̃A, it follows that L(Ã, µ) is equal to logarithm of the
norm of the greatest eigenvalue of any representative of ˆ̃A.
Furthermore, Proposition 3.1.4 also have a counterpart for PSL(2,R) cocycles. In
order to state it, recall that a sequence {L̃n}n in PSL(2,R) is said to converge
to L̃ ∈ PSL(2,R) if there are representatives L and Ln in SL(2,R) of L̃ and L̃n,
respectively, so that the sequence {Ln}n converges to L in SL(2,R).
Proposition 3.1.5. Let L̃n ∈ PSL(2,R) be a sequence converging to [[id]] and, for
each n ∈ N let ηn be an L̃n-invariant measure on P1 converging to 12(δp + δq) for some
3.1 Preliminary results 21
-
p, q ∈ P1 with p 6= q. Then for every n sufficiently large either L̃n is hyperbolic or
L̃n = [[id]].
This result follows easily from Proposition 3.1.4: for every L̃n ∈ PSL(2,R) we can
take a representative of L̃n in SL(2,R) with positive trace and apply the aforemen-
tioned result to these representatives.
3.2 Proof of Theorem A
Let f : M → M , A : M → SL(2,R) and µ be given as in Theorem A and suppose
there exists a sequence {Ak}k∈N in Hα(M) with λu(Ak, µ) = λs(Ak, µ) = 0 for
every k ∈ N and such that Ak converges to A in Hα(M).
For each k ∈ N, let mk be an ergodic FAk -invariant probability measure on M × P1
projecting on µ where FAk is defined similarly to FA. The set M(µ) of probabil-
ity measures on M × P1 that project down to µ is sequentially compact (see [24,
Lemma 6.4]). Hence, passing to a subsequence if necessary, we may assume that
the sequence {mk}k converges in the weak* topology to some measure m which is,
as one can easily check, FA-invariant and projects on µ. In order to prove Theorem
A we are going to analyze these families of measures and its respective disintegra-
tions.
3.2.1 Continuity and convergence of conditional measures
Recall that WsL = {(x, y) ∈ M ×M ;x ∼sL y} is compact and, since (x, y,A) → H
s,Axy
is continuous, given a sequence {Ak}k∈N converging to A in Hα(M),
{WsL ∋ (x, y) → Hs,Akxy }k∈N
is equi-continuous for k sufficiently large. Then, using this, Proposition 3.1.1 and
its proof in [3] we have:
Corollary 3.2.1. For every k sufficiently large there exists an su-invariant disintegra-
tion {mkx : x ∈ M} of mk with respect to the partition {{x} ×P1 : x ∈ M} and µ such
that
{M ∋ x → mkx}k≫0 is equi-continuous.
As an application of this corollary we get that
Proposition 3.2.2. The measure m is su-invariant and admits a continuous disinte-
gration {mx}x∈M with respect to {{x}×P1}x∈M and µ so thatmkx converges uniformly
on M to mx.
In order to prove the previous proposition we need the following auxiliary result.
22 Chapter 3 Partially hyperbolic base dynamics
-
Lemma 3.2.3. Let X and Y be compact metric spaces, µ a Borel probability measure
on X and {νk}k∈N be a sequence of probability measures on X × Y projecting on µ
and converging in the weak* topology to some measure ν. Then for every measurable
function ρ : X → R in L1(µ) and every continuous function ϕ : Y → R,
limk→∞
∫
ρ× ϕdνk =∫
ρ× ϕdν.
Proof. Let ϕ be a continuous real valued function on Y . It is well-known that
the continuous functions are dense in L1(µ) (see for instance [21, Appendix A.5]).
Thus, given ε > 0, we can take ρ̂ : X → R a continuous function so that∫
X|ρ̂− ρ|dµ <
ε
2 supϕ.
By the weak convergence, we can take k0 ∈ N such that for every k > k0,
∣
∣
∣
∣
∫
ρ̂× ϕdνk −∫
ρ̂× ϕdν
∣
∣
∣
∣
<ε
2.
Then, for k > k0,
∣
∣
∣
∣
∫
ρ× ϕdνk −∫
ρ× ϕdν
∣
∣
∣
∣
< supϕ
∫
X|ρ̂− ρ|dµ+
∣
∣
∣
∣
∫
ρ̂× ϕdνk −∫
ρ̂× ϕdν
∣
∣
∣
∣
< ε.
We now proceed to the proof of Proposition 3.2.2:
Proof of Proposition 3.2.2. For each k ∈ N, let {mkx}x∈M be the disintegration of
mk given by Corollary 3.2.1. We start observing that by the equicontinuity of the
disintegrations and Arzelà-Ascoli’s theorem, for every continuous function ϕ : P1 →
R, there exists a subsequence of{
∫
P1ϕdmkx
}
ksuch that
∫
P1ϕdm
kjx → Ix(ϕ)
uniformly on M . Moreover, fixing ϕ, the uniform convergence implies that Ix(ϕ) is
continuous in M .
Taking a dense subset {ϕj}j∈N of the space C0(P1) of continuous functions and
using a diagonal argument, passing to a subsequence if necessary, we can suppose
that∫
P1ϕdmkx → Ix(ϕ)
for every ϕ ∈ C0(P1). It is easy to see that Ix defines a positive linear functional on
C0(P1). Consequently, by Riesz-Markov’s theorem, for every x ∈ M there exists a
measure m̂x on P1 such that Ix(ϕ) =∫
ϕdm̂x.
3.2 Proof of Theorem A 23
-
On the other hand, letting {mx}x∈M be a disintegration of m with respect to {{x}×
P1}x∈M and µ and invoking Lemma 3.2.3 it follows that for every continuous func-
tion ϕ : P1 → R and any µ-positive measure subset D ⊂ M ,∫
D
∫
P1ϕdmkxdµ =
∫
M×P1χD × ϕdmk →
∫
M×P1χD × ϕdm =
∫
D
∫
P1ϕdmxdµ.
Consequently, mx = m̂x for µ almost every x ∈ M . Thus, extending mx = m̂xfor every x ∈ M we get a continuous disintegration of m such that mkx → mxuniformly on x ∈ M . In particular, by the equicontinuity of the holonomies and the
su-invariance of mk for every k it follows that m is also su-invariant as claimed.
From now on we work exclusively with the disintegrations {mkx}x∈M and {mx}x∈Mof mk and m, respectively, given by Corollary 3.2.1 and the previous proposition.
Recall we are assuming λu(A,µ) > 0 > λs(A,µ) hence, by the Oseledets’ theorem
there exist a decomposition R2 = Es,Ax ⊕ Eu,Ax . Moreover, we have the following
result, for its proof we refer the reader to [4, Proposition 3.1]:
Lemma 3.2.4. If λu(A,µ) > 0 > λs(A,µ), every FA-invariant probability measure
m that projects down to µ may be written as a convex combination ams + bmu with
a, b ≥ 0 and a + b = 1, where m∗ is an FA-invariant measure projecting on µ such
that its disintegration {m∗x}x∈M with respect to µ satisfies m∗x(E
∗x) = 1 for ∗ ∈ {s, u}.
It follows from the lemma above that for any FA-invariant measure m, its condi-
tional measures are of the form mx = aδEu,Ax + bδEs,Ax for some a, b ∈ [0, 1] such
that a + b = 1 where here and in what follows we abuse notation and identify a 1-
dimensional linear space E with its class [E] in P1. Furthermore, Avila, Santamaria,
Viana in [3, Theorem D] established that under the hypothesis of Theorem A every
su-invariant section of a continuous fiber bundle π : E → M is continuous in M .
Thus we have:
Lemma 3.2.5. There exist continuous and su-invariant functions which coincide with
x → Es,Ax , Eu,Ax for µ-almost every point. By su-invariance we mean that for every
(admissible) choice of x, y, z ∈ M , Hs,Axy E∗x = E
∗y and H
u,Axz E
∗x = E
∗z for ∗ ∈ {s, u}.
Proof. Recallmk is a FAk -invariant measure such thatmk → m. Since λu(Ak, µ) = 0
for every k ∈ N we get that∫
ΦAkdmk = 0 where ΦAk : M × P1 → R is given by
ΦAk(x, v) = log‖Ak(x)v‖
‖v‖ . On the other hand,
∫
ΦAkdmk →∫
ΦAdm.
24 Chapter 3 Partially hyperbolic base dynamics
-
Thus,∫
ΦAdm = 0 which implies that
aλ+(A,µ) + bλ−(A,µ) =∫
ΦAdm = 0.
Furthermore, since A(x) ∈ SL(2,R) for all x ∈ M we have λ+(A,µ) = −λ−(A,µ).
Therefore, a = b = 1/2. Now, by Proposition 3.2.2 we know that {mx}x is su-
invariant. Consequently, since Eu,Ax is u-invariant and Es,Ax is s-invariant, it follows
δEu,Ax
=1
a(mx − bδEs,Ax
)
is also s-invariant. Analogously, Es,Ax is u-invariant. In particular, Eu,Ax and E
s,Ax are
su-invariant. Continuity follows easily from [3, Theorem D] as mention above.
From now on we think of Es,Ax and Eu,Ax as continuous functions defined for every
x ∈ M .
3.2.2 Excluding the atomic case with a bounded number of atoms
In this subsection we prove thatmkxk can not have a bounded number of atoms (with
bound independent of k) for infinitely many values of k ∈ N and any xk ∈ M . The
general case will be reduced to this one. In order to do so, we need the following
lemma.
Lemma 3.2.6. If mky has an atom for some y ∈ M , then there exists j = j(k) ∈ N
such that for every x ∈ M , there exist v1x, . . . vjx ∈ P
1 so that
mkx =1
j
j∑
i=1
δvix .
Proof. Let vy ∈ P1 be such thatmky(vy) = β > 0 and for every x ∈ M , let γx be an su-
path joining y and x. Taking wx = HAkγx vy, by the su-invariance of the disintegration
{mkx}k it follows that mkx(wx) = β for every x ∈ M . Thus, considering
L = {(x, vx) ∈ M × P1; mkx(vx) = β}
we get that it is FAk -invariant. Indeed,
mkf(x)(PA(x)vx) = (PA(x))∗ mkx(PA(x)vx) = m
kx(vx) = β
Consequently, since mk is ergodic and
mk(L) =
∫
mkx
(
L ∩(
{x} × P1))
dµ ≥ β > 0.
3.2 Proof of Theorem A 25
-
it follows that mk(L) = 1. In particular, mkx(
L ∩(
{x} × P1))
= 1 for µ-almost every
x ∈ M . Otherwise, if the set D ⊂ M where mkx(
L ∩(
{x} × P1))
< 1 has positive
measure, then
mk(L) =
∫
Mmkx
(
L ∩(
{x} × P1))
dµ(x)
=
∫
Dmkx
(
L ∩(
{x} × P1))
dµ(x) +
∫
Dcmkx
(
L ∩(
{x} × P1))
dµ(x)
<
∫
Ddµ(x) +
∫
Dcdµ(x) = 1.
By the definition of L, this implies that
mkx =1
j
j∑
i=1
δvix ,
where 1j = β (in particular, j does not depend on x).
This is true for every x in a full measure set M̂ ⊂ M . To prove that this claim holds
true for every x ∈ M , given x ∈ M take some su-path γx from a point z ∈ M̂ to x.
By su-invariance, if we define ωix = HAγxv
iz we have
mkx =(
HAγx
)
∗mkz =
1
j
j∑
i=1
δωix .
We now proceed to prove that for infinitely many k non conditional measure of mkcan have a bounded number of atoms with bound independent of k. The proof is
going to be by contradiction. Assume there exist a subsequence k’s such that there
exist x ∈ M such that mkx has a bounded number of atoms.
So, passing to a subsequence and using the previous lemma suppose mkx has j(k)
atoms and that the sequence {j(k)}k is bounded. Restricting again to a subse-
quence, if necessary, we may assume that j(k) is constant equal to some j ∈ N. In
particular, since
mx =1
2δ
Es,Ax+
1
2δ
Eu,Ax,
for k sufficiently large mkx has an even number of atoms. Thus, writing
mkx =1
j
j∑
i=1
δvik
(x)
and reordering if necessary we may suppose that vik(x) → Eu,Ax for i ≤
j2 and
vℓk(x) → Es,Ax for ℓ >
j2 . Moreover, by Proposition 3.2.2, such convergence is
uniform.
26 Chapter 3 Partially hyperbolic base dynamics
-
Observe now that for each k there exists some xk ∈ M such that Ak(xk)vikk (xk) =
vjkk (f(xk)) for some ik ≤j2 and jk >
j2 . Otherwise, the set
L =⋃
x∈M{x} × {v1k(x), . . . v
j2
k (x)}
would be FAk -invariant with measure
mk(L) =
∫
mkx({v1k(x), . . . v
j2
k (x)})dµ =1
2,
contradicting the ergodicity. Thus, restricting to a subsequence, if necessary, we
may assume without loss of generality that vikk (xk) = v1k(xk) and v
jkk (xk) = v
jk(xk)
for every k ∈ N and that xk → x. In particular,
A(x)Eu,Ax = limk→∞
Ak(xk)v1k(xk) = lim
k→∞vjk(f(xk)) = E
s,Af(x),
a contradiction.
Summarizing, we can not have a subsequence {ki}i so that the sequence {j(ki)}i is
bounded where j(k) stands for the number of atoms of mkx (which is independent
of x ∈ M).
3.2.3 Conclusion of the proof
Now we consider the general case. The idea of the proof is to use Proposition 3.1.5
to reduce the proof to the case presented in Section 3.2.2,
Given x ∈ M let γ be a non-trivial su-loop at x. In particular, from Lemma 3.2.5 it
follows that
HAγ E∗,Ax = E
∗,Ax
for ∗ ∈ {s, u}. Consequently, either HAγ is hyperbolic or HAγ = ± id.
If there exist x and γ such that HAγ is hyperbolic then, by the convergence of
HAkγ to HAγ , it follows that H
Akγ is also hyperbolic for every k ≫ 0. Thus, since
(
HAkγ
)
∗mkx = m
kx, it follows that m
kx is atomic and has at most two atoms for ev-
ery k ≫ 0 but from Section 3.2.2 we know this is not possible. So, we get that
HAγ = ± id for every su-loop at x and every x ∈ M and therefore HÃγ = [[id]] for
every su-loop at x and every x ∈ M where à is defined in 3.1.3.
From Proposition 3.1.5 we get that either there exists a non-trivial su-loop γ at
some point x ∈ M and a sequence {kj}j going to infinite as j → +∞ so that HÃkjγ
is hyperbolic for every j and thus HAkjγ is also hyperbolic for every j, or HÃkγ = [[id]]
for every su-loop γ and every k > kγ for some kγ ∈ N.
3.2 Proof of Theorem A 27
-
Arguing as we did above we conclude that the first case can not happen. So, all we
have to analyze is the case when HÃkγ = [[id]] for every su-loop γ and every k > kγfor some kγ ∈ N.
If there exists k0 ∈ N so that kγ ≤ k0 for every su-loop γ then HÃkγ = [[id]] for all
k > k0 and for all γ. Making the change of coordinates given in 3.1 for every k > k0(recall Section 3.1.3) we get the that L(Ãk, µ) is equal to the logarithm of the norm
of the greatest eigenvalue of any representative of ˆ̃Ak(x), whereˆ̃Ak(x) is a constant
element of PSL(2,R), and ˆ̃Ak(x) →ˆ̃A(x). In particular,
λu(Ak, µ) = L(Ãk, µ)k→+∞−−−−→ L(Ã, µ) = λu(A,µ)
which is a contradiction.
Recall that in order to perform the change of coordinates in 3.1 it is enough to
assume that HÃkγ = [[id]] for every (K′, L′)-loop γ for some K ′, L′ > 0. To conclude
the proof of Theorem A, in view of the previous argument, we only have to show
that we can not have kγ arbitrarily large for (K ′, L′)-loops.
Let kγ be minimum for its defining property, that is, HÃkγ = [[id]] for every k > kγ
and HÃkγγ 6= [[id]]. Suppose that for each j ∈ N there exist xj ∈ M and a (K ′, L′)-
loop γj at xj so that kγjj→+∞−−−−→ +∞. Passing to a subsequence we may assume
xjj→+∞−−−−→ x and γj
j→+∞−−−−→ γ where γ is an su-loop at x. This can be done because
each γj has at most K ′ legs and each of them with length at most L′. In particular,
if γj is defined by the sequence xj = zj0, z
j1, . . . , z
jnj = xj then nj ≤ K
′ for every j.
Thus, passing to a subsequence we may assume nj = n ≤ K ′ for every j ∈ N and
zjij→+∞−−−−→ xi for every i = 1, . . . , n, consequently γ is the su-loop defined by the
sequence x = x0, x1, . . . , xn = x.
Now, since HÃγ = [[id]], HÃkγjγj
j→+∞−−−−→ HÃγ and H
Ãkγjγj 6= [[id]] it follows from
Proposition 3.1.5 (recall Proposition 3.2.2) that HÃkγjγj is hyperbolic for every j ≫ 0
and thus HAkγjγj is also hyperbolic for every j ≫ 0. Consequently, m
kγjx is atomic
and has at most two atoms for every x ∈ M and every j ∈ N which again from
Section 3.2.2 we know is not possible concluding the proof of Theorem A.
Remark 3.2.7. We observe that Theorem A can also be proved using the techniques of
couplings and energy developed in [5]. We chose to present the previous proof because
it is shorter and also different. It is also worth noticing that a similar result was
obtained by Liang, Marín and Yang [20, Theorem 6.1] for the derivative cocycle under
the additional assumption that f has a pinching hyperbolic periodic point. In our
context, such a hypothesis would immediately imply that all the conditional measures
mkx are atomic with at most two atoms for every k ≫ 0. In particular, Theorem A
would follow from the results of Section 3.2.2.
28 Chapter 3 Partially hyperbolic base dynamics
-
3.3 Examples
At this section we present two examples of fiber-bunched cocycles with nonvanish-
ing Lyapunov exponents over a partially hyperbolic map which are accumulated by
cocycles with zero Lyapunov exponents. The construction of this examples is based
on an example constructed by Wang and You in [26, Theorem 2]. Let us present
this result.
Let X be a Cr, r ≥ 1, compact manifold, T : X → X be ergodic with a normalized
invariant measure ν and A(x) be a SL(2,R)-valued function on X. We say that the
dynamical system (x, v) 7→ (T (x), A(x)v) in X × R2 is a quasi periodic cocycle if the
base of the system is a rotation on the torus. That is, if X = Tm = Rm \Zm and
T = Tω : x 7→ x+ ω.
Let ω be an irrational number. We call a rational number p/q a best approximation
of ω if every other rational fraction with the same or smaller denominator differs
from ω by a greater amount. In other words if the inequalities 0 < s ≤ q, and
p/q 6= r/s imply that∣
∣
∣
∣
ω −p
q
∣
∣
∣
∣
<
∣
∣
∣
∣
ω −r
s
∣
∣
∣
∣
.
Irrational numbers have an unique continued fraction expansion:
ω = [a0 : a1, a2...] = a0 +1
a1 +1
a2 + · · ·
where a0 ∈ Z and an ∈ N for n ≥ 1. The continued fraction expansion of an
irrational number is infinite, whereas rational numbers have finite but not unique
continued fraction expansions. For the proof of the properties of the continued
fraction expansion that we mention here we refer to [18].
Given an irrational number ω = [a0 : a1, a2, ...] the n-th convergent is the rational
number pn/qn = [a0 : a1, ..., an]. The sequence (pn/qn)n converges to ω and this
sequence is the best approximation to ω: for any other rational number a/b such
that b ≤ qn we have∣
∣
∣
∣
ω −pnqn
∣
∣
∣
∣
<
∣
∣
∣
∣
ω −a
b
∣
∣
∣
∣
.
Moreover, we say ω is of bounded type if there exist M > 0 such that an ≤ M for all
n.
With this definitions we are able to state Wang and You’s result.
Theorem 3.3.1. [26, Theorem 1] Consider quasi-periodic SL(2,R) cocycles over S1
with ω being a fixed irrational number of bounded-type. For any 0 ≤ l ≤ ∞, there
3.3 Examples 29
-
exists a cocycle Al ∈ C l(S1, SL(2,R)) with arbitrarily large Lyapunov exponent and
a sequence of cocycles Ak ∈ Cl(S1, SL(2,R)) with zero Lyapunov exponent such that
Ak → Al in the C l topology. As a consequence, the set of SL(2,R)-cocycles with
positive Lyapunov exponent is not C l open.
3.3.1 Proof of Theorem B
Let ω be an irrational number of bounded type and f0 : S1 → S1 be given by
f0(t) = t+2πω where S1 is the unit circle. Let A0 : S1 → SL(2,R) be the cocycle Ar
given by Theorem 3.3.1. And {Ak}k be a sequence in Cr(S1,SL(2,R)) converging
to A0 in the Cr topology, so that λu(Ak, ν) = 0 for every k ∈ N where ν denotes the
Lebesgue measure on S1 (also given by Theorem 3.3.1) .
Now, given f1 : N → N , a volume-preserving Anosov diffeomorphism of a compact
manifold N , let us consider the map f : M := S1 ×N → M given by
f(t, x) = (f0(t), f1(x))
and let  : M → SL(2,R) be given by
Â(t, x) = A0(t).
Thus, defining Âk(t, x) = Ak(t) and denoting by µ the Lebesgue measure on M we
have that the limit of Âk is Â. Moreover,
λu(Âk, µ) = λu(Ak, ν) = 0
for every k ∈ N and
λu(Â, µ) = λu(A0, ν) > 0.
Consequently, since f is a volume-preserving partially hyperbolic and center-bunched
diffeomorphism and f1 may be chosen so that (Â, f) is fiber-bunched, we complete
the proof of Theorem B.
3.3.2 Random product cocycles
We now present another construction showing that given any real number λ > 0, we
have a fiber-bunched cocycleA over a partially hyperbolic and center-bunched map
f so that λu(A,µ) = λ which can be approximated by cocycles with zero Lyapunov
exponents. We start with a general construction.
Let Σ = {1, . . . , k}Z be the space of bilateral sequences with k symbols and σ : Σ →
Σ be the left shift map. Given maps fj : K → K and Aj : K → SL(2,R) for
30 Chapter 3 Partially hyperbolic base dynamics
-
j = 1, . . . , k where K is a compact manifold, let us consider f : Σ × K → Σ × K
and A : Σ ×K → SL(2,R) given, respectively, by
f(x, t) = (σ(x), fx0(t)) and A(x, t) = Ax0(t).
The random product of the cocycles {(Aj , fj)}kj=1 is then defined as the cocycle over
f which is generated by A. Observe that this definition generalizes the notion of
random products of matrices explaining our terminology. Indeed, taking K as being
a single point we recover the aforementioned notion.
Differently from the case of random products of matrices where one have continuity
of Lyapunov exponents (see [5],[9], [24]), in the setting of random products of
cocycles Lyapunov exponents can be very ‘wild’. This is what we exploit to construct
our next example.
Let f0 : S1 → S1 and ν be as in the previous example and let A0 ∈ Cr(S1, SL(2,R))
be given by Theorem 3.3.1 so that λu(A0, ν) > λ.
Taking f1 : S1 → S1 to be f1(t) = t and A1 : S1 → SL(2,R) given by A1(t) = id,
let (A, f) be the random product of the cocycles (A0, f0) and (A1, f1) as defined
above. Thus, letting η be the Bernoulli measure on Σ defined by the probability
vector (p0, p1) where p0 is so that p0λu(A0, ν) = λ and considering µ = η × ν, we
are going to prove that the cocycle generated by A over f has positive Lyapunov
exponents and is accumulated by cocycles with zero Lyapunov exponents.
Indeed, let {A0,k}k be a sequence in Cr(S1, SL(2,R)) converging to A0 for which
the cocycle (A0,k, f0) satisfies λu(A0,k, ν) = 0 for every k ∈ N whose existence is
guaranteed by our choice ofA0 and Theorem 3.3.1. Also let {A1,k}k be the sequence
such that A1,k = id for every k ∈ N and (Ak, f) be the random product of (A0,k, f0)
and (A1,k, f1). It is easily to see that Akk→∞−−−→ A.
Now, for µ-almost every (x, t) ∈ Σ × S1,
λu(Ak, µ, x, t) = limn→∞
1
nlog ‖Ank(x, t)‖.
Thus, observing that Ank(x, t) = Aτn(x)0,k (t) where
τn(x) = #{
1 ≤ j ≤ n; σj(x)0 = 0}
,
it follows that
λu(Ak, µ, x, t) = limn→∞
τn(x)
n
1
τn(x)log
∥
∥
∥Aτn(x)0,k (t)
∥
∥
∥ = p0λu(A0,k, ν).
3.3 Examples 31
-
In particular, λu(Ak, µ, x, t) is constant equal to λu(Ak, µ) for µ-almost every (x, t) ∈
Σ × S1. Analogously, λu(A,µ) = p0λu(A0, ν). Consequently,
λu(Ak, µ) = 0 for every k ∈ N and λu(A,µ) = λ > 0
as claimed. Observe that despite the fact of not being smooth, the map f is partially
hyperbolic in the sense of the expansion and contraction properties when Σ is en-
dowed with the usual metric. Moreover, it is center-bunched and the cocycle A is
fiber-bunched.
32 Chapter 3 Partially hyperbolic base dynamics
-
4Probability distributions with noncompact support
In this chapter we present our results regarding the continuity of Lyapunov expo-
nents when the base measure has non compact support. We first analyze the semi-
contiuity relative to the weak* topology and to the Wasserstein topology in Section
4.1. Finally, in Section 4.2, we prove that the Wasserstein topology is not enough to
guarantee the continuity of the Lyapunov exponents.
4.1 Semicontinuity
Let us begin by giving an example that shows that if we only assume convergence
in the weak* topology, the Lyapunov exponents are not necessary semicontinuous.
On the other hand, in Section 4.1.2 we use the convergence of the first moments
provided by the Wassertein topology, to prove the semicontinuity relative to this
topology.
4.1.1 Semicontinuity counterexample with weak* topology
At this section we proof Theorem 2.4.3. Thus, we construct a sequence of measures
qk converging to q in the weak* topology such that λ+(qk) ≥ 1 while λ+(q) = 0.
We begin by defining the function α : N → SL(2,R) by
α(2k − 1) =
(
σk 0
0 σ−1k
)
α(2k) =
(
σ−1k 0
0 σk
)
where (σk)k is an increasing sequence such that σ1 > 1 and σk → +∞.
Let µ = qZ be a measure in M where q is the measure on SL(2,R) given by
q =∑
k∈Npkδα(k),
with∑
pk = 1, 0 < pk < 1 for all k ∈ N.
33
-
The key idea to construct this example is to find pk and σk such that log ‖A‖ ∈ L1(µ)
and satisfying the hypothesis above. Consider 0 < r < 1/2 < s < 1, and l = s/r > 1.
Let us take σk = elk
for all k, which is an increasing sequence provided that l > 1.
For k ≥ 2 take p2k−1 = p2k = rk. Since 0 < r < 1/2 it is easy to see that
∑
k≥3pk = 2
∑
k≥2p2k = 2
∑
k≥2rk = 2
r2
1 − r< 1
We have to choose p1 and p2 such that∑
pk = 1. Then, it is enough to take
p1 = p2 =1
2
(
1 − 2r2
1 − r
)
.
We continue by showing that log ‖A‖ ∈ L1(µ). This is an easy computation,
∫
Mlog ‖A‖dµ = 2p2 log σ1 + 2
∑
k≥2p2k log σk = 2p2l + 2
∑
k≥2sk.
Since 0 < s < 1 this geometric series is convergent. Moreover, since p2k−1 = p2k for
all k then λ+(q) = 0.
What is left is to construct the sequence qn. Fix n0 > 1 large enough so 12(
1 − 2 r2
1−r
)
>
l−n for all n ≥ n0, and consider qn =∑
k qnk δα(k) where for n ≥ n0
qn2n = l−n + rn,
qn2 =1
2
(
1 −r2
1 − r
)
− l−n
qnk = pk other case.
Thus, since qnk converges to pk when n goes to infinite for all k, it is easy to see that
qn converges in the weak* topology to q.
The proof is completed by showing that λ+(qn) ≥ 1 for n large enough. It follows
easily since,
λ+(qn) = |qn2n−1 − q
n2n| log σn + |q
n1 − q
n2 | log σ1 = l
−nln + l−n+1,
which is equal to 1 + l−n+1 ≥ 1 for all n ≥ n0.
4.1.2 Semicontinuity relative to the Wasserstein topology
We now consider the Wasserstein topology in P1(SL(2,R)). The advantage of using
this topology is that all probability measures in P1(SL(2,R)) have finite moment
of order 1. Therefore, the Lyapunov exponents always exist. This observation is a
34 Chapter 4 Probability distributions with non compact support
-
direct consequence of the fact that log : [1,∞) → R is a 1-Lipschitz function and,
‖α‖ ≥ 1 for every matrix α ∈ SL(2,R), because
∫
log ‖A(x)‖dµ =∫
log ‖α‖dp ≤∫
d(α, id)dp < ∞.
Before beginning the proof of Theorem 2.4.4 we need to recall some important re-
sults regarding the relationship between Lyapunov exponents and stationary mea-
sures.
A probability measure η on P1 is called a p-stationary if
η(E) =
∫
η(α−1E)dp(α),
for every measurable set E ∈ P1 and α−1E = {[α−1v] : [v] ∈ E}.
Roughly speaking, the following result shows that the set of stationary measures for
a measure p is closed for the weak* topology.
Proposition 4.1.1. Let (pk)k be probability measures in SL(2,R) converging to p in
the weak* topology. For each k, let ηk be pk-stationary measures and ηk converges to
η in the weak* topology. Then η is a stationary measure for p.
Furthermore, when f is the shift map, it is well-known that
λ+(p) = max
{∫
Φdp× η : η p− stationary}
, (4.1)
where Φ : SL(2,R) × P1 → R is given by
Φ(α, [v]) = log‖αv‖
‖v‖.
For more details see for example [24, Proposition 6.7].
Let us begin the proof of Theorem 2.4.4. In order to do this we will prove that λ+(p)
is upper semi-continuous. The case of λ−(p) is analogous.
Let (pk)k be a sequence in the Wasserstein space P1(M) converging to p, that is
W (pk, p) → 0. For each k ∈ N let ηk a stationary measure that realizes the maximum
in (4.1). That is:
λ+(pk) =
∫
Φdpkdηk.
Since P1 is compact then M(P1) the set of all invariant measures in P1 is sequen-
tially compact (see [21, Proposition 2.1.6]). Then, passing to a subsequence if nec-
essary, we can suppose ηk converges in the weak* topology to a measure η which,
as established by Proposition 4.1.1, is a p-stationary measure.
4.1 Semicontinuity 35
-
Let ǫ > 0, we want to prove that there exist a constant k0 ∈ N such that for each
k > k0∣
∣
∣
∣
∫
Φdpkdηk −∫
Φdpdη
∣
∣
∣
∣
< ǫ.
In order to do this we need to consider some properties of the Wasserstein topology.
First of all, since the first moment of p is finite, there exist K1 a compact set of
SL(2,R) such that∫
Kc1
d(α, id)dp <ǫ
36. (4.2)
Moreover, by Proposition 2.3.1, since pkW
−→ p there exist R′ > 0 satisfying
lim supk
∫
d(α,id)>R′d(α, id)dpk <
ǫ
36,
then, there exist k′ > 0 such that for every k > k′
∫
d(α,id)>R′d(α, id)dpk <
ǫ
36. (4.3)
Take R > 0 big enough so B(id, R′) ∪ K1 ⊂ B(id, R) and define the compact set
K = B̄(id, R).
Since the function log : [1,∞) → R is 1-Lipschitz and ‖α‖ ≥ 1 for all α ∈ SL(2,R),
then
|Φ(α, [v])| =
∣
∣
∣
∣
log‖αv‖
‖v‖
∣
∣
∣
∣
≤ log ‖α‖ ≤ |‖α‖ − ‖ id ‖| ≤ d(α, id). (4.4)
Our proof starts with the observation that
∣
∣
∣
∣
∫
Φdpkdηk −∫
Φdpdη
∣
∣
∣
∣
≤
∣
∣
∣
∣
∫
K×P1Φdpkdηk −
∫
K×P1Φdpdη
∣
∣
∣
∣
+
∣
∣
∣
∣
∫
Kc×P1Φdpkdηk
∣
∣
∣
∣
+
∣
∣
∣
∣
∫
Kc×P1Φdpdη
∣
∣
∣
∣
.
On account of (4.3) it follows that∣
∣
∣
∣
∫
Kc×P1Φdpkdηk
∣
∣
∣
∣
≤∫
Kcd(α, id)dpk <
ǫ
3. (4.5)
Furthermore, (4.2) implies that
∣
∣
∣
∣
∫
Kc×P1Φdpdη
∣
∣
∣
∣
≤∫
Kcd(α, id)dp <
ǫ
3. (4.6)
36 Chapter 4 Probability distributions with non compact support
-
We now proceed to analyze the integral:
∣
∣
∣
∣
∫
K×P1Φdpkdηk −
∫
K×P1Φdpdη
∣
∣
∣
∣
≤
∣
∣
∣
∣
∫
K×P1Φdpkdηk −
∫
K×P1Φdpkdη
∣
∣
∣
∣
+
∣
∣
∣
∣
∫
K×P1Φdpkdη −
∫
K×P1Φdpdη
∣
∣
∣
∣
.
Consider ΦK = Φ|K×P1 the restriction of Φ to the compact space K ×P1. Then, ΦK
is uniformly continuous with the product metric. Hence, there exist δ = δ(ǫ) such
that for every [v] ∈ P1 and every α, β ∈ K satisfying d(α, β) < δ we have
|ΦK(α, [v]) − ΦK(β, [v])| <ǫ
18.
Moreover, by the compactness of the set K we can find α1, ..., αN ∈ K such that
K ⊂ ∪Ni=1B(αi, δ). Therefore, the convergence of (ηk)k to η in the weak* topology
implies that for each i = 1, ..., N there exist ki > 0 such that for k > ki∣
∣
∣
∣
∫
P1ΦK(αi, [v])dηk −
∫
P1ΦK(αi, [v])dη
∣
∣
∣
∣
<ǫ
18.
Take k′′ = max{k1, ..., kN }. From the above it follows that given α ∈ K there exist
i su